Predicting Bandwidth Utilization on Network Links Using Machine Learning

Maxime Labonne, Charalampos Chatzinakis, Alexis Olivereau

TL;DR

Network congestion usually triggers reactive fixes, but predicting it beforehand can prevent packet loss and latency altogether. This paper introduces a machine learning approach using Long Short-Term Memory (LSTM) networks to predict network link bandwidth utilization 15 seconds into the future with under 3% error. The authors also demonstrate how this predictive model can be integrated with a Software-Defined Networking (SDN) controller to proactively route traffic and avoid bottlenecks before they occur.

What Problem Does This Paper Address?

Modern networks rely on reactive mechanisms to handle congestion. Whether it is allocating more bandwidth, prioritizing traffic via Quality of Service (QoS) rules, or load balancing across multiple paths, these actions are typically taken after a bottleneck is detected. By the time these systems react, users are already experiencing packet loss, buffering, and increased latency.

The authors tackle the challenge of proactive network management. By accurately predicting future bandwidth utilization on network links, network administrators (or automated SDN controllers) can reroute traffic or apply rate limits before a link reaches capacity, thereby maintaining a high Quality of Service.

Methods

The authors built a simulated network using VirtualBox and generated randomized TCP/UDP traffic using iPerf3. To capture the network's state, they used two standard Linux tools: Netmate (for network flow statistics) and Dstat (for system resource metrics like CPU and RAM).

  • Feature Engineering: Because Netmate generates data per active flow, the authors aggregated this data using statistical functions (min, max, mean, sum, count) to create 86 link-level features. These were combined with 29 Dstat features for a total of 115 features.
  • Modeling: They framed this as a supervised time-series forecasting problem, aiming to predict the future_bitrate 15 seconds ahead (using a 3-second data collection interval and a 5-step offset). They compared three models:
  • ARIMA (a traditional time-series baseline using only 1 feature: historical bitrate).
  • MLP (a Multilayer Perceptron feedforward neural network using all 115 features).
  • LSTM (a recurrent neural network using all 115 features).
  • Real-Time SDN Integration: Finally, they deployed the trained LSTM in a real-time Software-Defined Networking setup using Open vSwitch. The SDN controller received telemetry every 3 seconds, predicted traffic 15 seconds out, and proactively installed flow rules to drop or load-balance traffic if predicted utilization exceeded specific thresholds.

Datasets

The authors generated a custom simulated dataset rather than using an existing benchmark.

  • The dataset was created over a three-day simulation of a multi-router topology.
  • It contains around 85,000 samples per network interface, totaling 2 million samples (12 GB) across 22 monitored interfaces.
  • The data includes randomized flow parameters (bitrate, duration, inter-arrival time) to simulate varied traffic conditions ranging from zero traffic to maximum link capacity (100 Mbps).

Key Insights and Ideas

  • Memory matters for network traffic: The LSTM significantly outperformed the MLP. Because network traffic has complex temporal dependencies and bursty patterns, the LSTM's inherent ability to "remember" past variations is crucial for accurate forecasting.
  • Single-feature forecasting is insufficient: The ARIMA model, which only looked at historical bandwidth usage, failed to produce usable results. This highlights that predicting network congestion requires a holistic view of the network state (packet sizes, inter-arrival times, system resources), not just past bandwidth numbers.
  • Proactive SDN is highly viable: The authors successfully closed the loop by feeding the LSTM's predictions directly into an SDN controller. This proves that ML-driven forecasting can be practically applied in real-time to trigger load-balancing before a link hits 100% capacity, effectively smoothing out network performance.

Key Results

  • LSTM Dominance: The LSTM model achieved highly accurate predictions, with an average Root Mean Squared Error (RMSE) of 0.0122 and a Mean Absolute Error (MAE) of 0.0042.
  • Error Margins: The LSTM's prediction error rarely exceeded 3% (equivalent to an error of just +/- 300 Kbps on their 100 Mbps simulated links). In contrast, the MLP's error often exceeded 20%, and ARIMA's error hit 40%.
  • Real-time Application: In the SDN experiment, the controller successfully used the LSTM to predict bandwidth 15 seconds into the future. When predicted utilization crossed a threshold (e.g., 50% or 80% capacity), the controller proactively and successfully spread traffic across other available links or dropped incoming packets to prevent overload.

Limitations and Future Work

  • Feature Bloat: The authors acknowledge that not all 115 features (especially some system-level Dstat metrics) are likely necessary. Future work will explore dimensionality reduction to lower the computational cost and the bandwidth overhead required to send telemetry to the SDN controller.
  • Simulation vs. Reality: The current dataset relies on iPerf3, which struggles to generate truly realistic, bursty network traffic. They plan to test the system on real hardware using actual application traffic (HTTP, FTP, RTP).
  • Predicting System Depletion: Beyond just bandwidth, the authors suggest predicting future CPU and RAM usage on routers, as resource exhaustion on the hardware level also leads to packet loss and latency.

Summary updated

Abstract

Predicting the bandwidth utilization on network links can be extremely useful for detecting congestion in order to correct them before they occur. In this paper, we present a solution to predict the bandwidth utilization between different network links with a very high accuracy. A simulated network is created to collect data related to the performance of the network links on every interface. These data are processed and expanded with feature engineering in order to create a training set. We evaluate and compare three types of machine learning algorithms, namely ARIMA (AutoRegressive Integrated Moving Average), MLP (Multi Layer Perceptron) and LSTM (Long Short-Term Memory), in order to predict the future bandwidth consumption. The LSTM outperforms ARIMA and MLP with very accurate predictions, rarely exceeding a 3% error (40% for ARIMA and 20% for the MLP). We then show that the proposed solution can be used in real time with a reaction managed by a Software-Defined Networking (SDN) platform.
BibTeX
@inproceedings{labonne2020predicting,
  author = {Maxime Labonne and Charalampos Chatzinakis and Alexis Olivereau},
  title = {{Predicting Bandwidth Utilization on Network Links Using Machine Learning}},
  booktitle = {EuCNC},
  year = {2020},
  url = {https://openreview.net/forum?id=01dBN9zCMJ},
}

Loading related content...

Figure